Generalizing Over Uncertain Dynamics for Online Trajectory Generation

We present an algorithm which learns an online trajectory generator that can generalize over varying and uncertain dynamics. When the dynamics is certain, the algorithm generalizes across model parameters. When the dynamics is partially observable, the algorithm generalizes across different observations. To do this, we employ recent advances in supervised imitation learning to learn a trajectory generator from a set of example trajectories computed by a trajectory optimizer. In experiments in two simulated domains, it finds solutions that are nearly as good as, and sometimes better than, those obtained by calling the trajectory optimizer on line. The online execution time is dramatically decreased, and the off-line training time is reasonable.